ChemOntoRule Offers Verifiable AI for Chemistry Problem Solving

While large language models (LLMs) can interpret natural-language chemistry questions, their internal reasoning processes are often opaque, making it difficult to inspect, constrain, or validate their solutions. This paper introduces ChemOntoRule, a proof-of-concept symbolic core designed to provide verifiable AI assistance for school-level chemistry problems. The core innovation is a task-centric ontology, built specifically around the concepts, properties, relations, and procedures required for a defined set of chemistry problems, rather than aiming for a universal representation. This ontology is combined with deterministic Python rules covering fundamental chemistry patterns like electronic structure and oxidation states. A fallback mechanism handles problem types not yet covered by general rules. Tested on 300 human-authored and validated chemistry problems, the system achieved a 98.67% match rate with reference answers. This high performance, particularly from the ontology-driven rule subset, suggests a future architecture where LLMs primarily translate user queries into a structured ontological task frame, allowing a verifiable symbolic core to perform the actual reasoning.